TLDRocket
27 July 2026
The shape of AI investment is shifting away from raw compute and toward efficiency and data. Multiverse Computing's €500 million Series C round at a €2 billion valuation signals serious institutional confidence in model compression—the art of making AI systems smaller, faster, and cheaper without sacrificing capability. The Spanish startup's CompactifAI technology is becoming the infrastructure layer that companies deploying models actually need, as the era of throwing more parameters at problems runs into both cost and practical constraints. That's one storyline. The other is about physical AI's hidden bottleneck: the data itself. Encord is tackling a brutal economics problem—annotated training data for robotics costs twenty times more to produce than basic video footage, yet is worth a hundred times more. The company's experiments with brain wave sensors and muscle electrical signals suggest the next frontier in humanoid robot training isn't better algorithms but smarter ways to generate the dense, real-world manipulation data that currently doesn't exist at scale. Together, these stories point toward a maturing AI market where competitive advantage lives not in model size but in compression efficiency and data scarcity. The next wave of AI builders won't be the ones with the biggest models—they'll be the ones who can afford to run them, and who can actually afford to train them properly.
Read the full briefing →
